#6272: Teaching Audio Models to Hear Beyond Words
Speech recognition rewards ignoring background noise, but audio-language models train specifically on environmental sounds and temporal scenes.
#6271: Why Video Search Often Works Better as Text
A one-minute video compressed into a single vector undergoes roughly five orders of magnitude of raw data reduction.
#6267: Why Text-to-Speech Fails at Explaining Code
A study comparing Google TTS, ChatGPT, Gemini, ElevenLabs, and Copilot found every system failed to communicate indentation, scope, and code comments.
#6266: Balancing Speech Training Data With Square Roots
A fine-tuning recording app weighs tasks by the square root of unread sentences so massive datasets don't crowd out smaller prompt sets.
#6265: Why Upgrading Embedding Models Demands Reindexing
Cohere's Embed Five Fast queries against a Pro index reach 98.4 percent of performance without regenerating stored document vectors.
#6264: Choosing Omni Embedding Models for Mixed Media
Jina's omni model trains only 0.35 percent of its weights while aligning with existing text indices.
#6263: Why Numeric Data Lacks a Default Embedding Model
Popular time-series models like IBM's Granite TSPulse forecast trends rather than provide canonical vectors for similarity search.
#6257: Where Your Prompts Actually Go
Selecting US-only inference on Anthropic's direct API incurs a 1.1x multiplier across token categories, while global routing is the standard rate.
#6255: Routing Claude Code Tasks to Local Ollama Models
A feature request to support local subagent endpoints in Claude Code was closed for inactivity after less than two months.
#6254: Where a Synthetic Voice Gets Its Accent
A neural speech interface maps motor cortex signals into phonemes drawn from an eclectic training set of human voice recordings.
#6253: Why Context Windows Stalled at a Million Tokens
A dense 70B model requires about 328 gigabytes of KV cache to serve a single one-million-token prompt, exceeding the memory of an Nvidia B300 chip.
#6245: Why Inference Engines Change Benchmark Scores
A study found a 16.6 percentage point spread on GSM8K across different inference engines running identical weights on the same GPU.
#6244: When Your Fine-Tune Isn't Done Learning Yet
Falling loss, 90% word error rate, and a model that still says "Derry gym lag." What undertraining actually looks like — and when more epochs won't...
#6243: Inside the Mixture of Experts Router
What actually decides which expert handles your token? A look inside the router, from sigmoid affinities to load balancing.
#6242: Training a Model on Your Own Taste
What does it actually take to train a small model on your editorial taste using pairwise preferences and DPO?
#6241: Why Your Synthetic Data Isn't as Diverse as You Think
Fifty-nine percent of one GPT-4 dataset started with "Once upon a time." Correct examples can still make a terrible training set.
#6240: Gold vs Silver Training Data: What Makes Labels Trustworthy
Human-checked "gold" labels and machine-made "silver" ones aren't as different as the names suggest — and the gap matters.
#6239: Testing Synthetic Data Generators Before You Scale
A benchmark trains student models on each generator's output to find out which model actually writes useful training data.
#6238: Which AI Models Actually Generate Good Training Data?
Open weights vs closed APIs for synthetic data, and why IFEval scores don't tell you which model will hold up.
#6237: Small Models, Big Questions: When Cheap Fables Get Expensive
Three million fables for pennies — but can a small model tell a good example from a plausible one?